Server energy saving system and method based on a hysteresis compensation type dynamic thermal impedance model

Abstract

The application discloses a server energy-saving system and method based on a lag compensation type dynamic thermal impedance model, relates to the field of server energy saving, and solves the fan asthma phenomenon of a large amplitude sine fluctuation of a fan PWM duty cycle under constant computing power load through the following steps: collecting operation data of a server hardware bottom layer in real time, constructing a multi-dimensional heat-computing power coupling time sequence graph, solving a heat conduction time constant under a current environment of hardware based on the multi-dimensional heat-computing power coupling time sequence graph, establishing a lag compensation type dynamic thermal impedance model based on the heat conduction time constant, constructing a thermodynamic phase space according to the operation data, analyzing a noise reduction evolution track in combination with the lag compensation type dynamic thermal impedance model, performing a delay-induced strong resonance analysis based on the noise reduction evolution track, generating a weighted heat resonance state vector, activating an adaptive gain MPC predictive damping scheduling strategy based on the weighted heat resonance state vector, constructing and solving a dynamic weighted multi-objective optimization function, and generating an optimal damping scheduling instruction.

Description

Technical Field

[0001] This invention relates to the field of server energy saving, specifically to a server energy saving system and method based on a hysteresis-compensated dynamic thermal impedance model. Background Technology

[0002] In high-performance computing task scheduling scenarios, even if the server is performing steady-state tasks such as matrix operations with a constant instruction set strength factor, theoretically, the computing power density and heat dissipation should maintain a dynamic balance. However, in actual operation, a pathological oscillation phenomenon known as "fan asthma" often occurs. Specifically, despite a constant computing load, the fan PWM duty cycle exhibits large, sinusoidal fluctuations resembling breathing, accompanied by significant noise fluctuations. This oscillation not only causes severe acoustic pollution but also leads to the cooling system operating in a non-steady-state, high-energy-consumption range for extended periods. Frequent fan acceleration and deceleration consume a large amount of additional mechanical power, increasing the server's non-computing energy consumption and severely reducing the overall energy efficiency of the data center. The root cause is that existing scheduling systems mostly rely on reading the values ​​of digital temperature sensors distributed on the CPU package surface or calling the platform environment control interface. These data are essentially the apparent temperature after the core heat flow penetrates the primary thermal interface material and the heat sink, and is low-pass filtered by the physical medium. They cannot map the instantaneous power changes in the active junction region of the wafer in real time, resulting in a physical spatial transmission lag in the temperature field sensed by the system.

[0003] This perception lag caused by the physical medium leads to a severe misalignment between the millisecond-level control loop frequency of the scheduling algorithm and the second-level heat conduction time constant of the underlying hardware. When the system triggers overheat protection actions such as task migration or frequency reduction based on the lagging surface temperature, the core temperature has already dropped excessively due to thermal inertia; and when the system senses the low temperature and releases computing power again, the core temperature will overshoot due to the existence of the pure lag dead zone. This phase lag caused by the mismatch between physical thermal inertia and logical scheduling response speed causes the system to form a closed limit loop trajectory with its head and tail connected in the thermodynamic phase space, falling into a delay-induced strong resonance state. This causes the server to be in a tug-of-war between overheating and frequency reduction and performance release under constant computing power demand, forming the so-called fan asthma, and unable to converge to the steady state of heat-computing power coupling. This results in the server being in a tug-of-war between overheating and frequency reduction and performance release under constant computing power demand, unable to converge to the optimal energy-efficient operating point of heat-computing power coupling, causing a large amount of computing power waste and power consumption, making it difficult to meet the stringent requirements of green computing for energy saving and consumption reduction.

[0004] To address the aforementioned shortcomings, a technical solution is provided. Summary of the Invention

[0005] To address the technical problems raised in the background section, this invention is proposed. Embodiments of this invention provide a server energy-saving system and method based on a hysteresis-compensated dynamic thermal impedance model.

[0006] The objective of this invention can be achieved through the following technical solution: a server energy-saving method based on a hysteresis-compensated dynamic thermal impedance model, comprising the following steps:

[0007] Real-time acquisition of server hardware underlying operation data, mapping of the operation data to a unified time axis, and construction of a multi-dimensional thermal-computing power coupling time series graph;

[0008] Based on the multidimensional thermal-computing coupling time series diagram, the heat conduction time constant under the current hardware environment is calculated, and a hysteresis-compensated dynamic thermal impedance model is established based on the heat conduction time constant.

[0009] A thermodynamic phase space is constructed based on the operating data. The noise reduction evolution trajectory is analyzed in conjunction with the hysteresis compensation dynamic thermal impedance model. Delay-induced strong resonance analysis is performed based on the noise reduction evolution trajectory to generate a weighted thermal resonance state vector and switch the MPC predictive damping scheduling mode.

[0010] After switching to the MPC predictive damping scheduling mode, a dynamic weighted multi-objective optimization function is constructed and solved based on the weighted thermal resonance state vector to generate the optimal damping scheduling command.

[0011] Furthermore, the analysis steps for constructing the multidimensional heat-computing power coupling time series map are as follows:

[0012] The core real temperature field and fan PWM duty cycle are obtained through the in-band sensor interface, and the instruction set intensity factor and cache miss rate are extracted simultaneously through the operating system performance counter to form a heterogeneous multidimensional state flow.

[0013] Sliding window timing analysis is performed on the heterogeneous multidimensional state flow. The heat dissipation flux is calculated based on the spatial gradient of the core real temperature field. The computing power power density per unit time is deduced by combining the instruction set strength factor and cache miss rate, and a heat-power-heat dissipation coupled feature vector is generated.

[0014] The thermal-power-heat dissipation coupling feature vector is mapped into a topological structure using a time-series graph neural network, where nodes represent the physical core and edge weights represent the thermal conduction correlation strength, generating a multidimensional thermal-computing power coupling time-series graph.

[0015] Furthermore, the analysis steps for the core true temperature field are as follows:

[0016] A three-dimensional heat conduction geometry is constructed based on physical packaging parameters, and the discrete sampling data of the sensor is mapped to the boundary thermal potential energy constraint of the geometry surface to construct a reverse heat flow cone pointing towards the chip core.

[0017] An isothermal manifold surface cluster is generated within the reverse heat flow cone. The path backtracking is performed using the geometric axiom that the heat flow gradient vector is orthogonal to the isothermal surface, and the core true temperature field is obtained by superimposing the energy flux conservation constraint.

[0018] Furthermore, the analytical steps for establishing the hysteresis-compensated dynamic thermal impedance model are as follows:

[0019] Step response segments are extracted from the multidimensional thermal-computing coupling time series graph, and discrete transfer function coefficients are identified using the recursive least squares method with forgetting factor.

[0020] Based on the discrete transfer function coefficients, the heat conduction time constant, pure time delay dead zone parameter, and system steady-state gain are analyzed.

[0021] Based on the heat conduction time constant, the pure lag dead time parameter, and the system steady-state gain, a first-order inertial element model with pure lag is constructed, and a Kalman filter correction term is introduced to compensate for the ambient temperature drift, generating a lag-compensated dynamic thermal impedance model.

[0022] Furthermore, the analysis steps for generating the weighted thermal resonance state vector are as follows:

[0023] Using the pure lag dead time parameter and the heat conduction time constant as physical constraint boundaries, the oscillation frequency feature tensor is mapped to the lag-period coupling discriminant function to calculate the phase lag matching degree.

[0024] If the phase lag matching degree falls into the preset critical resonance range, the current state is determined to be a delay-induced strong resonance, a weighted thermal resonance state vector is generated, and the system switches to the MPC predictive damping scheduling mode. Otherwise, it is determined to be a transient load disturbance and the low-overhead PID feedback scheduling mode is activated.

[0025] Furthermore, the analysis steps for the oscillation frequency feature tensor are as follows:

[0026] Based on the limit cycle trajectory, the time regression period of the noise reduction evolution trajectory on the Poincaré section is extracted, and combined with the thermodynamic phase space dimension information, an oscillation frequency feature tensor is constructed.

[0027] Furthermore, the analysis steps for the limiting cycle trajectory are as follows:

[0028] A Poincaré section is set in the thermodynamic phase space, and the regression characteristics of the noise reduction evolution trajectory crossing the section are detected. If the Euclidean distance between the crossing points is less than a preset convergence threshold multiple times, the trajectory is determined to form a limit cycle trajectory and is confirmed to be in a thermal oscillation resonance state.

[0029] Furthermore, the analysis steps for the noise reduction evolution trajectory are as follows:

[0030] The collected instantaneous temperature data are subjected to first-order and second-order difference operations to extract dynamic feature vectors and construct a thermodynamic phase space.

[0031] The dynamic feature vector is mapped to discrete trajectory points in the thermodynamic phase space. The hysteresis-compensated dynamic thermal impedance model is used to eliminate non-periodic trajectory divergence. The discrete trajectory points are then corrected and reconstructed to obtain the noise-reduced evolution trajectory.

[0032] Furthermore, the analysis steps for generating the optimal damping scheduling instruction are as follows:

[0033] The second component in the weighted thermal resonance state vector is analyzed, and the prediction time domain length of MPC is dynamically adjusted based on the second component. The trajectory point corresponding to the current time is extracted from the denoised evolution trajectory, and the discrete state evolution equation is established with the trajectory point corresponding to the current time to obtain the initial prediction control decision matrix.

[0034] A dynamic weighted multi-objective optimization function is constructed using the predictive control decision matrix as the decision variable. The multi-objective optimization function includes a stability term that minimizes the variance of the temperature trajectory in the prediction time domain and a performance term that maximizes the computing power throughput. A fan noise penalty factor is added, and the stability term is dynamically weighted and amplified by the resonance suppression gain coefficient modulated by the weighted thermal resonance state vector.

[0035] The multi-objective optimization function is solved using a rolling time-domain optimization algorithm, and the optimal damping scheduling command is output under the premise of satisfying the core's highest safe temperature constraint.

[0036] A server energy-saving system based on a hysteresis-compensated dynamic thermal impedance model includes:

[0037] The time series graph module is used to collect the underlying operating data of the server hardware in real time, map the operating data to a unified time axis, and construct a multi-dimensional heat-computing power coupled time series graph;

[0038] Based on the multidimensional thermal-computational coupling time series graph, the model generation module calculates the heat conduction time constant under the current hardware environment and establishes a hysteresis-compensated dynamic thermal impedance model based on the heat conduction time constant.

[0039] The resonance analysis module constructs a thermodynamic phase space based on the operating data, analyzes the noise reduction evolution trajectory in combination with the hysteresis-compensated dynamic thermal impedance model, and performs delay-induced strong resonance analysis based on the noise reduction evolution trajectory to generate a weighted thermal resonance state vector and switch the MPC predictive damping scheduling mode.

[0040] After switching to the MPC predictive damping scheduling mode, the scheduling instruction module constructs and solves a dynamic weighted multi-objective optimization function based on the weighted thermal resonance state vector to generate the optimal damping scheduling instruction.

[0041] The independent edge inference module, independent of the server's main processor, serves as the physical execution carrier for the aforementioned modules, acquiring data and executing computational tasks via the sideband bus.

[0042] The lightweight model storage module stores the lightweight neural network parameters after pruning and quantization, providing parameter support for the time series graph module.

[0043] Compared with the prior art, the beneficial effects of the present invention are:

[0044] This invention collects real-time operational data from the server hardware, maps this data to a unified time axis, and constructs a multi-dimensional thermal-computing coupling time series graph. Based on this graph, it calculates the heat conduction time constant under the current hardware environment, establishes a hysteresis-compensated dynamic thermal impedance model, constructs a thermodynamic phase space based on the operational data, analyzes the noise reduction evolution trajectory using the hysteresis-compensated dynamic thermal impedance model, and performs delay-induced strong resonance analysis based on the noise reduction evolution trajectory to generate a weighted thermal resonance state vector. Through microsecond-level real-time hardware sensing, it effectively avoids the inherent conduction hysteresis of traditional temperature sensors due to crossing the thermal interface material and the physical medium of the heat spreader, fundamentally eliminating the phase delay dead zone in the control loop. It successfully breaks the closed limit loop trajectory formed by the misalignment of logical response and physical thermal inertia in the thermodynamic phase space, cutting off the generation path of delay-induced strong resonance. This completely solves the fan asthma phenomenon caused by large sinusoidal fluctuations in the fan PWM duty cycle under constant computing load, enabling the server's thermal-computing coupling system to quickly converge and lock into a globally optimal steady state. This not only eliminates unnecessary acoustic noise fluctuations and mechanical fatigue, but also significantly reduces the ineffective power consumption of the cooling system, achieving energy savings in terms of heat dissipation.

[0045] This invention, based on a weighted thermal resonance state vector, activates an adaptive gain MPC predictive damping scheduling strategy, constructs and solves a dynamically weighted multi-objective optimization function, generates optimal damping scheduling instructions, and establishes a preemptive scheduling model based on transient physical characteristics. This overcomes the hysteresis losses caused by repeated back-and-forth between overheating misjudgments and computing power release overshoot in traditional hysteresis feedback mechanisms. By achieving zero-phase synchronization between the task scheduling frequency and wafer-level instantaneous power mutations, the system can precisely squeeze the chip's thermal design power margin at the critical point without triggering hardware-forced frequency reduction, transforming the thermal capacity redundancy originally consumed in temperature oscillations into substantial effective computing power throughput. This not only avoids performance sawtooth fluctuations caused by sensor delays but also achieves deterministic improvement in floating-point performance per unit of energy consumption while ensuring hardware physical safety boundaries, achieving an optimal balance between computing power release and energy consumption, thus possessing energy-saving and efficiency-enhancing value. Attached Figure Description

[0046] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. The following drawings are not drawn to scale according to the actual size, but are intended to illustrate the main idea of ​​the present invention.

[0047] Figure 1 The flowchart shows a server energy-saving method based on a hysteresis-compensated dynamic thermal impedance model.

[0048] Figure 2 This is a system block diagram of a server energy-saving system based on a hysteresis-compensated dynamic thermal impedance model.

[0049] Figure 3 This is a schematic diagram of a reverse heat flow cone;

[0050] Figure 4 This is a schematic diagram of isothermal manifold and flux correction. Detailed Implementation

[0051] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are also within the scope of protection of the present invention.

[0052] Example 1

[0053] like Figure 1 As shown, the server energy-saving method based on the hysteresis-compensated dynamic thermal impedance model includes:

[0054] Step S1: Collect the underlying operating data of the server hardware in real time, map the operating data to a unified time axis, and construct a multi-dimensional heat-computing power coupling time series map;

[0055] The construction of the multidimensional heat-computing coupling time series map in step S1 specifically includes:

[0056] Step S101: Obtain the core real temperature field and fan PWM duty cycle through the in-band sensor interface, and simultaneously extract the instruction set intensity factor and cache miss rate through the operating system performance counter to form a heterogeneous multidimensional state flow;

[0057] Basic thermal data is acquired using an in-band sensor interface. Specifically, this is done through a hardware monitoring driver loaded at the operating system kernel level, which calls the Platform Environment Control Interface (PECI) or directly accesses the processor model's relevant registers (MSR) to read the raw sample values ​​of the digital temperature sensor (DTS) distributed on the CPU package surface using a high-frequency polling method. Steps S1011 and S1012 are executed to construct a three-dimensional heat conduction geometry and a reverse heat flow cone based on physical package parameters. By generating isothermal manifold surface clusters and combining heat flow path backtracking with energy flux conservation, the core true temperature field is obtained. Simultaneously, by reading the fan control register in the motherboard embedded controller, the ratio of the high-level duration to the signal period in the cooling fan drive signal at the current moment, i.e., the fan PWM duty cycle, is obtained. At the same time, the software layer calls... The operating system kernel-level performance monitoring unit driver accesses hardware performance counters to calculate the ratio of the number of high-energy-consuming instructions submitted to the total core clock cycles within the same sampling window. This calculates the instruction set strength factor, which characterizes the computational energy density. High-energy-consuming instructions include floating-point and vector instructions. The system also calculates the cache miss rate, which characterizes the data transfer overhead, by calculating the ratio of the number of access misses in the final cache (LLC) to the total number of cache references. Using the system's high-precision timestamp counter (TSC) as a unified time reference, the system aligns the core's actual temperature field and fan PWM duty cycle (physical dimension) with the instruction set strength factor and cache miss rate (logical dimension). After normalizing the data of different dimensions, the system encapsulates them into a time series, forming a heterogeneous multidimensional state flow that includes thermodynamic state, heat dissipation actions, and computational load characteristics.

[0058] The steps for analyzing the core true temperature field are as follows:

[0059] like Figure 3 As shown, step S1011: Construct a three-dimensional heat conduction geometry based on physical packaging parameters, and map the discrete sampling data of the sensor into the boundary thermal potential energy constraint of the geometry surface, and construct a reverse heat flow cone pointing towards the chip core;

[0060] like Figure 3 The diagram shown is a schematic of a reverse heat flow cone.

[0061] Reference Figure 3 The topmost strip with a grid texture represents the heat sink (IHS), with multiple black dots embedded within it indicating the measured sensors. The middle layer with a dotted texture represents the thermal interface (TIM). The rectangular area with diagonal lines at the bottom represents the chip core (Die). The inverted trapezoidal outline formed by the dashed lines around the perimeter represents the virtual computational domain boundary of the reverse heat flow cone. A node and connecting line in the middle of the diagram represent the micro-element nodes in the thermal resistance network. The thick solid line at the bottom represents the computational reference plane projected onto the bottom of the chip.

[0062] Physical packaging parameters are extracted from the CPU's physical design document. These parameters include the length and width dimensions of the packaging substrate, the spatial coordinates and dimensions of the die within the package, the thickness and thermal conductivity of the thermal interface material (TIM), and the thermal resistance characteristics of the vapor chamber (IHS). A three-dimensional Cartesian coordinate system is constructed with the geometric center of the top surface of the CPU package as the origin O. The direction parallel to the long side of the package is defined as the X-axis, the direction parallel to the wide side as the Y-axis, and the direction perpendicular to the package surface pointing towards the inner die as the Z-axis. Based on these physical packaging parameters, a multi-layered composite structure containing an IHS layer, a TIM layer, and a die layer is constructed in the coordinate system using a finite element mesh generation method; this is the three-dimensional thermal conductivity geometry. Discrete sampling data mapping is then performed, specifically by discretely sampling the temperature values ​​from the N sensors. As a known boundary condition, 'i' represents the sensor's number or serial number, based on the physical coordinates of each sensor in the circuit board design. Lock the mesh node corresponding to the surface of the 3D heat conduction geometry, and set the temperature scalar of that node to . These nodes with definite temperature values ​​are defined as boundary thermal potential energy constraint points. For blank areas on the surface of the geometry where no sensors are deployed, the Kriging interpolation algorithm is used to calculate the surface thermal potential energy distribution based on the constraint points, forming a continuous surface temperature boundary field. A reverse heat flow cone pointing towards the chip core is constructed. Specifically, a virtual computing domain is defined inside the three-dimensional heat conduction geometry. This computing domain is spatially represented as a frustum shape with the sensor distribution range on the package surface as the base and the physical projection area of ​​the CPU core (Die) as the top section. Within this cone, the solution direction of the heat conduction equation is set to the reverse time step or the reverse heat flow direction, that is, along the positive Z-axis from the surface to the inside. The reverse process of heat transfer from the core to the surface is reversed using the thermal resistance network method. A reverse heat flow cone model for inverting the true temperature field of the core is established, specifically including: spatial discretization of the reverse heat flow cone, along the... The axial direction divides the cone into The system consists of several layers, including a surface sensor layer, a middle heat spreader (IHS) layer, a thermal interface material (TIM) layer, and a bottom die layer. Each layer's micro-elements are defined as thermal resistance network nodes. The equivalent thermal resistance between nodes is calculated using Fourier's law of heat conduction and the formula... Calculate the thermal resistance between adjacent nodes , It refers to the characteristic length, which in this case means the conduction distance between two nodes or the thickness of a material layer. It is thermal conductivity, and A represents the cross-sectional area perpendicular to the heat flow direction, i.e., the common area of ​​contact between two nodes; establish a set of heat balance equations based on the conservation of node energy, and based on Kirchhoff's thermal circuit law, for any current node in the network. Establish The steady-state equations are given by summation over all equations relating to... Adjacent nodes , The current node Temperature value, Neighboring nodes Temperature value, It is a connection node With nodes The equivalent thermal resistance between them It is the summation symbol, representing summation over all integers. Connected The heat flux terms generated by the nodes are accumulated, and the equations of all nodes are solved simultaneously to construct the form: A linear system, in which The system's global thermal conductivity matrix is ​​composed of the inter-node thermal conductivity values, where the inter-node thermal conductivity values ​​are... , It is a global temperature column vector, containing an unknown global vector of temperatures for all nodes from the surface to the core. This is a heat source vector, representing the heat generation power applied to each node. It is allocated to the corresponding nodes based on the power distribution map of the chip design, using an area-weighted mapping method. Specifically, the geometric coordinates of each functional module in the chip layout are superimposed and aligned with the grid coordinates of the thermal resistance network. For each functional module, such as CPU cores, GPUs, and caches, the overlap area between each grid cell and each functional module is calculated. Assuming uniform power density within the functional module, the real-time power of each functional module is precisely projected and accumulated to the corresponding bottom core layer node based on its overlap area with the grid. The module power is discretized into local heat source values ​​for each node. For nodes in non-heat-generating layers, the heat source value is set to 0. Non-heat-generating layers include heat sinks and TIM materials. A reverse inference solution is then performed to obtain the global temperature column vector. Divided into known surface nodal temperature sub-vectors and the internal and core node temperature sub-vectors to be solved By analyzing the global thermal conductivity matrix By performing block processing and matrix inversion, the input is derived as the surface temperature distribution, and the output is... The linear transfer matrix model can quantitatively calculate the fine temperature field distribution of each layer inside the chip, especially the core layer, based on real-time changing surface temperature data.

[0063] For example, in the process of constructing the reverse heat flow cone and thermal resistance network, the size of the packaging substrate is set to be... The physical dimensions of a wafer die are Set the physical thickness of the thermal interface material (TIM) layer. Its thermal conductivity Physical thickness of the heat spreader IHS layer It is made of copper, and its thermal conductivity is... The reverse heat flow cone is discretized along the Z-axis as follows: There are 8 levels, with the IHS layer occupying 8 levels and the TIM layer occupying 2 levels. This is for two vertically adjacent micro-element nodes within the IHS layer. If the cross-sectional area of ​​the mesh (Right now Vertical spacing (Right now Then calculate the equivalent conductive thermal resistance between the nodes. Similarly, calculate the thermal resistance between nodes in the TIM layer. The constructed global thermal conductivity matrix Dimensions If surface sensor In coordinates The real-time sampled value at the location Then the temperature state of the corresponding boundary node is forcibly constrained to be By solving linear systems This allows us to invert the basic temperature distribution of the underlying core nodes after considering the differences in material thermal resistance.

[0064] Step S1011 establishes a reverse heat flow cone for the chip core. A primary prerequisite for resolving fan lag is obtaining the true heat source temperature, rather than the lagging surface temperature. Existing sensors are typically distributed on the CPU package surface or at specific points, while the actual heat source is the internal wafer. Step S1011 establishes the reverse heat flow cone to address the spatial heat conduction lag problem. The process involves constructing a three-dimensional geometry based on the CPU package substrate dimensions, wafer coordinates, TIM thickness, etc., mapping the discrete sensor data on the surface to boundary thermal potential constraints. It's called a "reverse heat flow cone" because the computational domain is a frustum-shaped cone projected from the bottom of the sensor distribution area on the package surface to the top section of the internal CPU core projection area. Within this cone, the thermal resistance network method is used to solve the heat conduction equation in reverse along the Z-axis (from the surface inwards). This step is the physical foundation of the entire technical solution. Without establishing this reverse cone, the system can only obtain surface temperature data buffered by the heat spreader and thermal conductive material. This data has already undergone low-pass filtering and delay in physics, making it unable to reflect instantaneous power surges in the core. Without this step, subsequent models will make predictions based on incorrect temperature inputs, and the scheduling algorithm will be perpetually processing past thermal states, failing to eliminate fan wheezing oscillations caused by sensing lag. Therefore, this step is absolutely indispensable for breaking through the physical thermal inertia barrier.

[0065] like Figure 4 As shown, step S1012: an isothermal manifold surface cluster is generated within the reverse heat flow cone, and the path backtracking is performed using the geometric axiom that the heat flow gradient vector is orthogonal to the isothermal surface, and the core true temperature field is obtained by superimposing the energy flux conservation constraint.

[0066] like Figure 4 The diagram shown is a schematic of isothermal manifold and flux correction.

[0067] Reference Figure 4 The wider area at the top of the diagram is the surface of the package, which is the starting surface for the discrete sampling data of the sensor. The nested dashed curves in the middle of the diagram represent the generated isothermal manifold surface clusters. The solid line that runs vertically downward through the isothermal surface at the center of the diagram is the heat flow back path. The arrow q in the diagram represents the heat flux density vector. The area at the bottom of the diagram that gradually narrows is the core layer of the chip, which is the core real temperature field after superposition calculation.

[0068] Specifically, refer to Figure 4The overall outline presents an inverted cone or frustum structure, wider at the top and narrower at the bottom. This is visually reflected by the "cross-sectional area (shrinking with depth)" label in the diagram. This structure corresponds to the microshunt mentioned in the specification, and its geometric shrinkage is controlled by the thermal diffusion angle of the packaging material. As the back-dive depth increases, i.e. along the Z-axis from the surface inward, the side length and cross-sectional area of ​​the shunt gradually decrease, vividly demonstrating the reverse process of heat diffusion from the tiny chip core to the huge heat sink. Figure 4 The dashed curves distributed in the middle layers represent isothermal manifold surface clusters. This is a continuous temperature potential energy surface constructed by the system based on discrete surface sensor data through global spatial interpolation. Figure 4 The solid line perpendicularly passing through these dashed lines represents the heat flow back path. This path is generated according to the geometric axiom that "the direction of heat flow is always strictly perpendicular to the isothermal surface." The algorithm starts at the surface sensor coordinates and performs a gradient descent search along the Z-axis, ensuring that the path is always orthogonal to each isothermal surface, thus accurately pinpointing the physical channels of heat conduction. Figure 4 The core's physical constraint lies in the conservation of energy flux. In the diagram, arrow q points towards the core. According to the law of conservation of energy, in a dielectric layer without an internal heat source, the total heat flux Q at any cross-section of the flow tube must remain constant. Figure 4 It can be seen that the cross-sectional area decreases significantly with increasing depth. To maintain a constant total heat flux Q, the heat flux density vector q must increase accordingly. This physical process, represented by the oscillation frequency characteristic tensor, is shown in the figure as follows: the heat flux is "compressed" into a smaller region, resulting in a much higher energy density at the core than at the surface. The system utilizes this principle to integrate and correct the heat flux along the backtracking path, ultimately achieving the desired result. Figure 4 At the core true temperature field at the bottom, the true temperature rise, after eliminating the influence of dielectric thermal resistance, was calculated, thus accurately locating the thermal conduction bottleneck inside the package.

[0069] Based on the discrete temperature data of each node in the constructed inverse heat flow cone, global spatial interpolation is performed. A three-dimensional interpolation algorithm maps the discrete grid node temperatures to a continuous temperature scalar field. Within this scalar field, all spatial points with the same temperature value are searched and connected, constructing a series of non-intersecting, nested, and smoothly transitioning surfaces in three-dimensional space, defined as an isothermal manifold surface cluster. Each surface represents a constant temperature potential energy surface. Based on scalar field gradient theory, the normal vector at any point on each surface in the isothermal manifold surface cluster is calculated, and this normal vector is then directed towards... The reverse vector of the quantity is defined as the heat flow gradient vector. Based on the geometric axiom in thermal physics that the direction of heat flow is always strictly perpendicular to the isothermal surface, the physical coordinates of the sensor on the surface of the package are used as the starting tracking point. A gradient descent path search is performed along the Z-axis from the surface inward. Specifically, a continuous three-dimensional curve that is always orthogonal to each isothermal surface traversed is constructed using the fourth-order Runge-Kutta numerical integration method, defined as the heat flow backtracking path. With the heat flow backtracking path as the central axis, a micro shunt tube with a cross-sectional area varying with the path depth is constructed, based on the thermal diffusion angle of the package material. The flow channel exhibits an inverted cone or truncated pyramid structure that gradually tapers from the package surface towards the chip core layer, and its geometric contraction follows a pattern that satisfies any backtracking depth. The side length of the flow tube at the location ,in Let be the initial side length of the surface element, whose value corresponds to the pixel physical size determined by the spatial resolution of the temperature acquisition device, and its cross-sectional area is . Based on Fourier's law of thermal conductivity, the measured temperature data of the package surface and the thermal conductivity of the material at that location are used. Calculate the heat flux density vector at the starting end, i.e., the surface, of the flow tube. ,in The temperature gradient is a continuous scalar temperature field in space, representing the rate of temperature change with location, rather than a simple numerical temperature value. The component perpendicular to the surface is determined in conjunction with surface convection heat transfer boundary conditions. Energy transfer constraints within the flow tube are set based on the law of conservation of energy, assuming the total heat flux across any cross-section of the flow tube in an encapsulated medium layer without an internal heat source. Keep constant and satisfy the relation ,in The heat flux density at this depth is due to the cross-sectional area. It exhibits a non-linear, monotonically decreasing trend with increasing depth, therefore It increases with depth; flux correction of the temperature field is performed, specifically by integrating the total heat flux along the heat flow back path, based on the thermal conductivity of each layer of material and the cross-sectional area of ​​the flow tube at different depths. The precise temperature rise from the surface to the core is calculated and superimposed onto a continuous temperature scalar field to eliminate interpolation errors. The temperature field superimposed from the chip core surface to the heat diffusion layer of the heat spreader is extracted to obtain the true temperature field of the chip core that conforms to the law of conservation of energy in physical fields. The chip core surface refers to the active junction region on the surface of the semiconductor wafer inside the package. This temperature field fully maps the dynamic picture of high-density heat flow from the microscopic silicon-based heat source through the primary thermal interface material to the macroscopic metal cover plate. Its core function is to accurately quantify the junction-shell thermal resistance and locate the thermal conduction bottleneck inside the package.

[0070] For example, when using energy flux conservation for temperature field correction, the thermal diffusion angle of the encapsulation material is set. Assume the initial side length of the heat flow back path on the surface element of the package is... (corresponding surface micro-element area) ), backtracking depth (Penetrating the IHS layer and TIM layer to reach the core surface). According to the geometric contraction law, the side length of the flow channel shrinks when reaching the chip core layer. The corresponding core micro-element area At this point, the area shrinkage ratio of the flow tube is approximately Times. If the heat flux density measured on the surface According to the law of conservation of energy Heat flux density at the core When performing flux correction, the average equivalent thermal conductivity along the path is assumed. The temperature rise correction is calculated by integrating the heat flow path. Under simplified calculations, if the total heat flow rate in the flow tube... And the total integral thermal resistance of the cone path The calculated core true temperature If the surface sensor reading ,but .this The temperature difference is the hysteresis temperature rise filtered out by the physical medium; the system will use... Instead As a benchmark for subsequent control.

[0071] Step S1012, within the geometric space constructed in S1011, further refines the temperature using energy flux conservation and geometric axioms. The direction of the heat flux gradient is always perpendicular to the isothermal manifold surface cluster. By starting from the surface sensor coordinates, tracing back along the heat flux gradient vector path, and combining this with the energy conservation law within the flow tube—that is, assuming the total heat flux of the flow tube cross-section is constant in a medium without an internal heat source—the increase in heat flux density is calculated using the nonlinear decrease in cross-sectional area with depth, thereby accurately calculating the true temperature rise of the core surface. This is a crucial step in obtaining the true value of the system; the root cause of fan asthma lies in the misalignment between the control loop frequency and physical thermal inertia. Without step S1012, the system cannot obtain the true core temperature field free from the influence of the physical medium's thermal resistance. Without this hysteresis-free true temperature as a reference, the identification of pure hysteresis time and thermal time constant in subsequent step S203 will lack an accurate reference frame, meaning it will be impossible to distinguish which delays are caused by physical transmission and which are measurement errors. Without it, the system cannot quantify the specific time difference of heat conduction from the core to the surface, and therefore cannot perform effective phase compensation in subsequent steps, and the oscillation problem cannot be solved at its root.

[0072] Step S102: Perform sliding window timing analysis on the heterogeneous multidimensional state flow, calculate the heat dissipation flux based on the spatial gradient of the core real temperature field, and deduce the computing power consumption density per unit time by combining the instruction set strength factor and cache miss rate, and generate a heat-power-heat dissipation coupled feature vector.

[0073] Set the sliding window length to And the sliding step size is The state flow is discretized into time slices. For the core true temperature field within each time window, which is defined as the superimposed temperature distribution covering the physical stack structure from the chip core die surface, through the primary thermal interface material TIM1 to the heat spreader IHS thermal diffusion layer, a three-dimensional difference algorithm is used to calculate the spatial gradient of the stack structure along the vertical heat flow direction. The vertical heat flow direction refers to the longitudinal path along the Z-axis, perpendicular to the horizontal surface of the chip, extending upwards from the die through TIM1 and pointing towards IHS; based on Fourier's law of thermal conductivity. Calculate heat flux ,in To determine the equivalent thermal conductivity of a multilayer composite structure, we provide a pre-determined physical thickness and inherent thermal conductivity of the chip core layer, TIM1 layer, and IHS layer, based on a series thermal resistance model formula. Calculate total thermal resistance The reciprocal is then taken and normalized to obtain the value, thus accurately characterizing the medium properties of heat conduction from the core to the outside. Indicates the first The inherent thermal conductivity of the layer material, Indicates the first The physical thickness of the layer material; the basic operational power consumption is obtained by monitoring the processor's current instruction execution frequency and multiplying it by the corresponding instruction set strength factor, and cache miss rate is also provided. As a power consumption weighting factor, a baseline threshold is set. Using a linear penalty function ,in The cache miss rate is mapped to a dimensionless correction coefficient greater than 1, using a preset energy consumption penalty gain constant. Multiply the base computing power by The corrected computing power-to-power density is obtained to reflect the additional energy consumption caused by pipeline waits due to missing data and high-frequency read / write operations of the memory controller; the calculated heat dissipation flux is also included. The corrected computing power power density and the current fan PWM duty cycle (as an active cooling boundary condition) are combined and normalized to generate a heat-power-cooling coupled feature vector, which reflects the dynamic balance of heat generation, heat conduction and heat removal. The current fan PWM duty cycle is used as the active cooling boundary condition. The normalization refers to the maximum value-minimum value method.

[0074] Step S103: Use a time-series graph neural network to map the thermal-power-heat dissipation coupling feature vector into a topological structure, where nodes represent physical cores and edge weights represent the thermal conduction correlation strength, generating a multidimensional thermal-computing power coupling time-series graph.

[0075] Construct a computing environment based on temporal graph neural networks, mapping the thermal-power-heat dissipation coupled feature vectors of multi-core processors to a graph topology. , where the set of nodes Each node in , N0 represents the total number of physical cores, uniquely representing a single physical core of the processor. A physical core is a hardware entity on the chip wafer with independent processing units, private cache, and independent thermal characteristics, rather than a logical thread at the operating system level. The heat dissipation flux generated in the previous steps... The thermal-power-heat dissipation coupling feature vector, composed of the corrected computing power power density and the current fan PWM duty cycle, is used as the input feature matrix. Load to the corresponding node Above, the PWM duty cycle is a globally fixed scalar at the current time step, but it is copied to each node as an environmental state feature. Define the edge set. The corresponding edge weights are used to characterize the strength of heat conduction correlation. Specifically, the method involves calculating the edge weights of any two physical cores. and Center distance in chip geometry , and These represent unique index identifiers for different nodes in the graph. This refers to the target core that is being observed during the current calculation process. Refers to the core of the target The neighboring core that generates heat conduction effects. Initial weights are calculated using the inverse distance decay function. The specific logic of this function is that the weight value is inversely proportional to the physical distance, that is, the distance between the two cores. The smaller the value, the higher the calculated weight. The larger the core, the easier it is for heat to escape from the core. Transmission to the core Conversely, the greater the distance, the less the weight decreases non-linearly, indicating a weak thermal effect. This weight... The physical difficulty of heat conduction, i.e., the strength of the heat conduction correlation, was precisely quantified. A temporal graph neural network model, consisting of a cascaded graph convolutional network (GCN) and a long short-term memory (LSTM) network, was run. In the spatial dimension, the graph convolutional layers, based on weights... For neighboring nodes Features are weighted and aggregated to nodes To simulate the lateral diffusion and mutual interference of heat flow between multiple cores, in the time dimension, LSTM units are used to capture the dynamic evolution of feature vectors with time step, and the node state vectors updated after spatiotemporal convolution operation are recombined to output a structured data set containing the thermal coupling topology between cores and the dynamic characteristics of time changes, that is, to generate the multidimensional thermal-computing coupling time series map.

[0076] Step S2: Based on the multidimensional thermal-computing coupling time series diagram, calculate the thermal conduction time constant under the current hardware environment, and establish a hysteresis-compensated dynamic thermal impedance model based on the thermal conduction time constant;

[0077] The establishment of the hysteresis-compensated dynamic thermal impedance model in step S2 specifically includes:

[0078] Step S201: Extract the step response segment from the multidimensional thermal-computing coupling time series graph and identify the discrete transfer function coefficients using the recursive least squares method with forgetting factor;

[0079] The node state data in the multidimensional thermal-computing power coupling time series graph are scanned in the time domain. Signal abrupt changes are determined by calculating the absolute value of the first-order difference between the input computing power power density or the fan PWM duty cycle signal at adjacent sampling times. When this absolute value exceeds a preset abrupt change threshold... When a step change occurs in the signal amplitude, that moment is marked as the step trigger moment. ,by Based on this, the time segment forward is... The time period is used as a preset length time period that includes the system steady-state baseline, where The value is determined based on the system's thermal time constant, typically taken as 3 to 5 times the sampling period to ensure that the variance of the temperature data within that time period is less than the preset background noise variance. Starting from the point of origin, temperature response data is monitored backwards. The sliding window method is used to calculate the slope of temperature change within the window. When continuous... The absolute value of the temperature change slope within each sliding window is less than the preset steady-state convergence threshold. When the temperature response reaches a new quasi-steady state, this moment is recorded as _____. Thus intercepting from arrive The continuous time series data is defined as a step response segment; for each physical core node, a discrete transfer function model describing its thermal dynamic characteristics is established, and the linear difference equation of this model is expressed as: ,in, This represents the sequence index at the current sampling time. Characterizing the first The actual core temperature output value at each sampling moment is extracted from the actual core temperature field of the chip constructed based on the law of conservation of energy in physical fields. Specifically, it is taken from the temperature data corresponding to the core node position in the temperature field obtained by cropping from the surface of the chip core to the heat diffusion layer of the heat spreader and superimposing the data. Characterizing the first The input energy at each sampling time, i.e., the corrected computing power power density. To output the lag order, To input the lag order, The summation index variable for the output items, with values ​​ranging from 1 to... , The summation index variable for the input items, with values ​​ranging from 1 to... , and These are the discrete transfer function coefficients to be identified; the coefficients are solved using the recursive least squares method with a forgetting factor. The specific process involves initializing the parameter estimation vector. and covariance matrix Set a forgetting factor Its value range is The forgetting factor This is used to adjust the algorithm's retention of historical data at each sampling time. Construct a regression vector containing historical input and output data. Calculate the gain vector ,in Let be the covariance matrix of the previous time step. This indicates a delay. Historical core temperature output value for each sampling period This indicates a delay. The historical input energy values ​​for each sampling period, where T represents the matrix transpose; calculate the prior prediction error at the current moment. ,in This is the parameter estimation vector from the previous time step; the parameter estimation value is obtained by correcting this error. And update the covariance matrix. ,in The identity matrix is ​​used; this process continues until all data points within the step response segment have been traversed, ultimately outputting the converged parameter estimation vector. Contained in As the coefficients of the identified discrete transfer function.

[0080] Step S202: Based on the discrete transfer function coefficients, analyze the heat conduction time constant, the pure time delay dead zone parameter, and the system steady-state gain;

[0081] Discrete transfer function coefficients Substituting into the Z-transform domain, constructing a system with input coefficients For numerator, output coefficient Let the discrete transfer function model with the denominator be the polynomial of the transfer function. Here is the characteristic equation, where For the complex frequency domain variables, solve the roots of the characteristic equation to obtain all the poles of the system, and select the real pole with the largest magnitude that lies inside the unit circle as the principal pole. The dominant pole determines the main dynamic response speed of the system. The heat conduction time constant is calculated using the mapping relationship between the dominant pole and the continuous domain time constant. The calculation formula is: ,in The sampling period of the system, Represents the natural logarithm operation. The physical meaning of this term represents the time required for the chip core to rise to 63.2% of its steady-state temperature under thermal excitation, reflecting the thermal inertia of the system and analyzing the input coefficient sequence. Determine the number of lag beats corresponding to the first non-zero significance coefficient in the sequence. The non-zero significance coefficient, i.e., the coefficient whose absolute value is greater than the preset noise threshold, is used to calculate the pure time delay dead zone parameter. The calculation formula is: ,in This characterizes the time delay between a change in computing power consumption and the detection of the temperature change by the temperature sensor. This delay mainly consists of the physical transfer time of heat passing through the silicon substrate and the thermal interface material (TIM). Based on the discrete transfer function model, let... The steady-state gain of the system is obtained by calculating the ratio of the sum of the numerator coefficients to the sum of the denominator coefficients. It characterizes the final temperature rise caused by unit power consumption, used to calibrate the steady-state amplitude of the thermal response in subsequent modeling, ensuring the accuracy of the model's predicted final temperature value. The output includes... The parameter set is used to complete the analysis of physical parameters.

[0082] Step S203: Based on the heat conduction time constant, the pure lag dead time parameter, and the system steady-state gain, construct a first-order inertial element model with pure lag, and introduce a Kalman filter correction term to compensate for ambient temperature drift, thereby generating a lag-compensated dynamic thermal impedance model.

[0083] Utilizing the thermal conduction time constant Pure time delay dead zone parameter and steady-state gain Construct a first-order inertial element model with pure time delay: ,in This represents the true temperature of the chip core. This value is extracted from the true temperature field of the chip core constructed based on the law of conservation of energy in physical fields. Specifically, it is taken from the temperature data corresponding to the core node location in the temperature field obtained by cropping and superimposing the temperature field from the chip core surface to the heat diffusion layer of the heat spreader. It was delayed The input power consumption over time, i.e., the corrected computing power power density, The environmental temperature disturbance term is represented by a variable that drifts slowly over time. To compensate for this temperature drift, the differential equation is discretized and transformed into a state-space equation, thus constructing a state vector. ,in, Indicates the first Real-time temperature status of the chip core. Indicates the first The ambient temperature state at any given time is modeled here as a random walk process. , For process noise, to simulate its slow drift over time, where This represents the current sampling time number in the Kalman filter iteration process, and the state prediction equation is constructed. ,in Indicates based on Information about time The prior estimate of the state at time step [time]. Indicates in The posterior estimate after Kalman filtering correction at time step [time]. Indicates in Before the moment The input power consumption value per cycle, i.e. the corrected computing power power density, reflects the hysteresis characteristic of heat conduction; Here is the state transition matrix. Both are input control matrices and are based on the sampling period. and time constant The equation is derived from the discretization of the continuity equation. The Kalman gain is calculated using the Kalman filter correction term. Using the original temperature sampling value Deviation from predicted temperature update status ,in, That is, the first The time-corrected state vector estimate includes the corrected core true temperature estimate. and the corrected ambient temperature estimate , Is the system in The raw temperature samples from the on-chip thermal sensor, read directly from the underlying hardware interface, are used as the observed true values ​​in the algorithm. These are compared with the predicted values ​​from the theoretical model to calculate the deviation and correct the system's state estimate, eliminating the cumulative error of the model over time. C is the observation matrix in the Kalman filter, used to map the system's internal state vector to the observation space. A hysteresis-compensated dynamic thermal impedance model is generated. This model refers to the entire recursive algorithm mentioned above, which includes state prediction and correction. Its core output logic lies in separating environmental influences; specifically, it involves using the core true temperature estimate from the state vector output by the Kalman filter. Subtract the estimated ambient temperature That is, calculating the pure temperature of the chip core. This yields a pure dynamic temperature rise curve that eliminates the influence of ambient temperature drift and is driven solely by chip power consumption. This calculation result is the output representation of the hysteresis-compensated dynamic thermal impedance model.

[0084] Step S203, based on the actual temperature obtained in S1012, constructs a model using Kalman filtering and a first-order inertial element. Its core principle is to abstract the physical heat conduction process into a mathematical expression of pure time delay plus thermal inertia, and to introduce ambient temperature as a random walk variable for separation. This is the bridge connecting the physical world and the control algorithm. Without this step, the scheduling system would not know the hardware's thermal response speed. The system must know that if the load is increased now, the core temperature will start to rise in milliseconds and how quickly it will reach steady state. Without this model, the MPC controller cannot predict future temperature trajectories and can only perform simple feedback control, which is the direct cause of the repeated fluctuations in overheat protection and performance release.

[0085] Step S204: Apply a preset step test load to the CPU in self-calibration mode, obtain the actual temperature response curve, compare the deviation between the actual temperature response curve and the hysteresis compensation dynamic thermal impedance model, and correct the parameters of the hysteresis compensation dynamic thermal impedance model.

[0086] The server's operating status is monitored in real time via a background monitoring service. When the server is detected to be in the startup phase before the operating system boots, or when the CPU utilization rate continuously falls below a preset idle threshold (e.g., 5%) for a preset time (e.g., 30 seconds), regular tasks are automatically suspended and a self-calibration mode is activated. In this mode, the scheduling system executes preset fixed-point arithmetic code to apply a calibrated step test load with a known power amplitude to the CPU core. The step power amplitude is denoted as [value missing]. The underlying temperature sensor is used to record the actual temperature response curve over time at a high-frequency sampling rate, and the actual steady-state temperature rise when the curve reaches thermal equilibrium is calculated and denoted as [missing value]. The criterion for reaching thermal equilibrium is that, within a preset sliding time window, for example, for 5 consecutive seconds, the absolute value of the rate of change of the sampled temperature data remains below a preset small threshold, such as 0.1℃ / s. This indicates that the temperature no longer rises significantly and physical thermal equilibrium has been reached. Simultaneously, to obtain a theoretical benchmark for comparison, the system uses the same step power amplitude as described above. As the input control variable in the first-order inertial element model of pure time delay. The values ​​are substituted into the differential equation of the constructed hysteresis-compensated dynamic thermal impedance model for calculation. At this point, the model structure and pure hysteresis time parameters are directly taken from the data in the previous steps, and only the currently stored parameters to be calibrated are used. and Calculate and output the theoretical pure temperature of the chip core. This involves generating a pure dynamic temperature rise curve and extracting the predicted steady-state temperature rise value when the pure dynamic temperature rise curve reaches a stable state. The criterion for determining the stable state is that when the increment of the temperature rise value output by the model within a continuous calculation step is less than the preset calculation accuracy tolerance, the mathematical model curve is considered to have converged to the theoretical final value. The system then calculates the actual steady-state temperature rise value. If the deviation rate from the predicted steady-state temperature rise exceeds a preset aging threshold, such as 10%, it is determined that aging of the thermal interface material or dust accumulation on the radiator has caused thermal resistance drift. Immediately, parameter reverse correction is performed. Specifically, the system steady-state gain is recalculated using the ratio of the actual measured steady-state temperature rise data to the input power, i.e., the updated system steady-state gain is... To calibrate the model's prediction accuracy of the final temperature rise, the heat conduction time constant is refitted based on the time taken for the actual temperature response curve to rise from the starting point to 63.2% of its actual steady-state temperature rise. Finally, the corrected version was used. and By overriding the original model parameters, the adaptive update of the hysteresis-compensated dynamic thermal impedance model is completed, greatly increasing the system's lifespan and robustness.

[0087] Step S3: Construct a thermodynamic phase space based on the operating data, analyze the noise reduction evolution trajectory in conjunction with the hysteresis-compensated dynamic thermal impedance model, and perform delay-induced strong resonance analysis based on the noise reduction evolution trajectory to generate a weighted thermal resonance state vector and switch the MPC predictive damping scheduling mode.

[0088] Step S3 specifically includes:

[0089] Step S301: Perform first-order and second-order difference operations on the collected instantaneous temperature data, extract the dynamic feature vector, and construct the thermodynamic phase space;

[0090] The instantaneous core temperature of the chip is obtained through a thermal sensor integrated inside the chip, at a preset sampling period. The actual temperature value of the chip during operation is collected in real time and recorded as follows: ,in This indicates the sampling time sequence, which has not yet eliminated the influence of environmental heat conduction and includes measurement noise. Based on the acquired raw temperature sequence. Numerical differencing is performed to calculate the dynamic characteristics of temperature changes, including calculating the first-order difference. , Indicates the acquisition chip Calculate the second-order difference based on the actual temperature value at time. , express The first-order difference of time. (The same time interval is used for...) The combination of instantaneous chip core temperature, temperature change rate, and temperature change acceleration defines the dynamic characteristic vector. Constructing the initial phase space involves establishing a three-dimensional coordinate system with the instantaneous chip core temperature as the X-axis, the temperature change rate as the Y-axis, and the temperature change acceleration as the Z-axis. This phase space differs from traditional time-domain waveform diagrams. By eliminating the explicit dependence of time variables, it uniquely represents the thermodynamic state of the system at any given moment as a coordinate point in space, thus providing a geometrical basis for subsequently identifying the system's periodic oscillations, i.e., limit cycles.

[0091] Step S302: Map the dynamic feature vector to discrete trajectory points in the thermodynamic phase space, use the hysteresis-compensated dynamic thermal impedance model to eliminate non-periodic trajectory divergence, and correct and reconstruct the discrete trajectory points to obtain the noise-reduced evolution trajectory.

[0092] Dynamic eigenvectors The discrete trajectory points are sequentially marked in the thermodynamic phase space. The lines connecting these points in space represent the original thermal evolution path, including environmental disturbances and noise. This relates to the pure temperature of the chip core. Performing the same second-order difference operation yields the theoretical rate of temperature change, which conforms to the laws of physical thermal inertia. and theoretical temperature change acceleration Construct a theoretical pure state vector In the thermodynamic phase space, the initial state vector of step S301 is... With the theoretical pure state vector For comparison, since ambient temperature changes are typically low-frequency and gradual, while measurement noise exhibits high-frequency and abrupt changes, the focus is on calculating the deviation between the two in the acceleration component. If the deviation exceeds the physical constraint tolerance set based on the chip's thermal capacity, then the deviation at that moment is determined to be... The discrete trajectory points are identified as non-periodic trajectory divergences that violate thermodynamic laws. The denoised evolution trajectory is obtained by correcting and reconstructing these discrete trajectory points. Specifically, a dynamic weighted fusion strategy based on the degree of deviation is employed, and the corrected dynamic eigenvector is obtained using a formula. Tuning of divergent noise points in phase space, where Defined as the confidence level correction coefficient, with a value range of Its value is dynamically determined by the degree to which the measurement point deviates from the thermodynamic manifold—when the monitoring data violates the chip's thermal inertia physical limit, i.e., is judged as divergent noise, then... Forced use of theoretically pure vectors that conform to deterministic physical laws Original measurement of replacement failure By utilizing the deterministic constraints of the physical model, outliers are pulled back to the true path of continuous evolution, thereby suppressing random noise and smoothly reconstructing the trajectory. When the deviation is within a reasonable range and is not a divergent noise point, some measurement information is retained. Taking a smaller value, specifically 0.1, will correct the dynamic eigenvector. By connecting them in phase space according to the time sequence, a noise reduction evolution trajectory that removes environmental temperature drift interference and conforms to the physical thermal response law is generated.

[0093] Step S303: Set a Poincaré cross section in the thermodynamic phase space and detect the regression characteristics of the noise reduction evolution trajectory crossing the cross section. If the Euclidean distance between the crossing points is less than the preset convergence threshold multiple times, it is determined that the trajectory forms a limit cycle trajectory and is confirmed as a thermal oscillation resonance state.

[0094] In the constructed three-dimensional thermodynamic phase space, a stable evolution interval of the noise reduction evolution trajectory is selected, and the average value of the tangent vectors of all trajectory points within this interval is calculated. This average value is defined as the average flow direction of the phase trajectory and used as the normal vector. Using the geometric center of this trajectory as a fixed point Construct a system that satisfies the equation The plane as the Poincaré section ,in Represents any coordinate point in phase space. Represents the vector dot product; discretizes the denoised evolution trajectory and monitors trajectory points. With cross section The relative position, when two consecutive sampling times and Corresponding trajectory points and satisfy and At that time, it is determined that the trajectory crosses the cross section in the positive direction, where The temporal index of the sampling point is used to calculate the precise coordinates of the crossing point through linear interpolation. And record the sequence of points that cross the cross section multiple times consecutively, and calculate the first... Second crossing point With the Second crossing point Euclidean distance in three-dimensional phase space ,in Indicates the counting index of the cross section. Denotes the L2 norm, if the distance If the convergence threshold is less than the preset convergence threshold, such as 1% of the phase space feature scale, then the trajectory is determined to form a closed limit loop trajectory, and is confirmed as a thermal oscillation resonance state.

[0095] Step S303, to accurately identify this specific oscillation, known as "fan asthma," rather than ordinary load fluctuations, introduces phase space reconstruction technology from nonlinear dynamics. A stable sinusoidal oscillation, in a three-dimensional phase space constructed using temperature, rate of temperature change, and acceleration of temperature change, will manifest as a closed limit cycle. By setting a Poincaré cross section, it detects whether the denoised evolution trajectory repeatedly crosses this cross section at similar positions. If the Euclidean distance between the crossing points converges, it indicates that the trajectory is closed, and the system has entered a stable thermal oscillation state. This is the core criterion for intelligent sensing; traditional threshold judgments cannot distinguish between a single high-temperature pulse and continuous thermal oscillations. Fan asthma is a systemic resonance that must be identified morphologically. Without step S303, the system cannot determine whether it is currently in an oscillation state and therefore does not know when to activate a special damping scheduling strategy. The system will misjudge the oscillation as a normal load change, continuing to incorrectly adjust the frequency, causing the oscillation to never stop. This step is the only means to distinguish between normal high load and pathological thermal oscillations.

[0096] Step S304: Based on the limit cycle trajectory, extract the time regression period of the noise reduction evolution trajectory on the Poincaré section, and construct the oscillation frequency feature tensor by combining the thermodynamic phase space dimension information;

[0097] Based on the limit cycle trajectory, extract the two consecutive positive crossings of the Poincaré section from this trajectory. The corresponding precise timestamps are denoted as follows: and Calculate the time difference between the two. This serves as the time regression period of the noise reduction evolution trajectory on the Poincaré section. Indicates the first The moment of the second crossing of the cross section Indicates the first The moment of the next cross-section; the fundamental frequency of thermal oscillation is calculated based on this period. Construct the oscillation frequency feature tensor ,in Characterizing the rate of thermal oscillation, Characterizing the periodicity of thermal state regression; this characteristic tensor As thermal fingerprint information, it can quantify the frequency attributes and dynamic complexity of the current thermal oscillation resonance state, which can be used for accurate matching of subsequent thermal fault modes.

[0098] Step S304, based on the limiting cycle detected in step S303, extracts the time difference of its crossing of the Poincaré cross section, calculates the regression period, and constructs a tensor containing the oscillation fundamental frequency. This is to quantify the fingerprint of the oscillation. Without this step, the system only knows that oscillation is occurring, but not how fast it is. Different oscillation frequencies correspond to different physical causes. Only by extracting precise frequency characteristics can they be compared with the physical time constant of the hardware in subsequent step S305 to confirm whether this is the resonance described in the technical problem.

[0099] Step S305: Using the pure lag dead time parameter and the heat conduction time constant as physical constraint boundaries, map the oscillation frequency feature tensor to the lag-period coupling discriminant function, and calculate the phase lag matching degree;

[0100] The standard pure time delay was obtained through an open-loop step response experiment. Compared with the standard thermal conduction time constant The system is determined to be in thermal equilibrium steady state, meaning that within a set time sliding window, the control quantity of the actuator remains constant, and the standard deviation of the temperature sensor sampling data is less than a preset steady-state noise threshold. This threshold is set based on the sensor accuracy and is used to eliminate environmental white noise interference. The applied amplitude is The step input signal is the control quantity of the actuator, such as the valve opening, which is adjusted instantaneously and maintained at an increased value. The system continuously records the response curve of the temperature sensor and calculates the temperature change in real time. The moment when the absolute value of the temperature change first exceeds three times the standard deviation of the steady-state noise is determined as the response moment. Time difference As a standard pure time delay dead zone This parameter characterizes the purely physical transmission delay caused by the fluid medium being transferred from the actuator to the sensor position; the time at which the temperature response curve rises to 63.2% of the total change is calculated. The time difference is defined as the standard heat conduction time constant. This parameter characterizes the dynamic response speed of the system in overcoming thermal inertia and transferring heat. The oscillation frequency characteristic tensor... Extract the time regression period. Based on the frequency domain analysis principle in control theory, a hysteresis-period coupling discriminant function is constructed. This function is used to calculate the total phase lag of the system during the current oscillation period. Its calculation formula is The first item The second term represents the pure time delay phase angle caused by the dead time. This represents the first-order inertial phase angle caused by thermal inertia. Let pi be the mathematical constant; based on this, the phase lag matching degree is defined. This is used to quantify the current total phase lag and the ideal self-excited oscillation phase condition, i.e., phase lag. The degree of matching of the radians is calculated using the following formula: ,in This indicates the absolute value of the match. The closer the value is to 1, the greater the current oscillation period. It is highly coupled with the physical hysteresis characteristics of the system, and is more consistent with the physical mechanism of internal self-excited oscillation.

[0101] Step S306: If the phase lag matching degree falls into the preset critical resonance range, the current state is determined to be a delay-induced strong resonance, a weighted thermal resonance state vector is generated, and the system is switched to the MPC predictive damping scheduling mode; otherwise, it is determined to be a transient load disturbance and the low-overhead PID feedback scheduling mode is activated.

[0102] The preset critical resonance range is determined based on historical operating data or offline simulation experiments. Specifically, this involves collecting multiple sets of sample data from the system under known delayed-induced strong resonance conditions (DET), where DET is a constant-amplitude oscillation state. The phase lag matching degree of each sample set is calculated, its distribution pattern is statistically analyzed, and the 95% confidence interval of this distribution is selected as the critical resonance interval. For example, the following is set... , This interval characterizes the tolerance range for the system to satisfy the self-excited oscillation phase condition, allowing for model linearization errors and sensor noise interference; the phase lag matching degree... Compare with that interval; if Falling into the critical resonance range If the current thermal oscillation is induced by internal delay and inertial feedback mechanisms, it indicates that the current state is a delay-induced strong resonance. The system immediately switches to the MPC predictive damping scheduling mode, which is specifically designed to handle complex resonance conditions. Switching to this mode aims to utilize the multi-objective optimization algorithm in subsequent step S4 to specifically address high-order nonlinear thermal oscillation problems by applying precise damping suppression through predicting future temperature trends. Simultaneously, the oscillation frequency feature tensor is weighted based on the matching degree to generate a weighted thermal resonance state vector. ,in This vector represents the fundamental frequency of thermal oscillation. The oscillation intensity reliability and frequency characteristics are combined and input into step S4 for further processing; conversely, if... If the temperature fluctuation does not fall within this range, it indicates that the current temperature fluctuation period does not match the system's physical hysteresis characteristics and lacks the physical basis for self-sustaining oscillation. This is determined to be a transient load disturbance caused by external random interference, activating the low-overhead PID feedback scheduling mode. In this state, since the system does not fall into a complex resonance closed loop, there is no need to use a high-computing-power prediction model. To avoid wasting computational resources in a non-resonant steady state, the system omits the complex matrix solution process and adopts a lightweight proportional-integral-derivative algorithm. This mode requires very little CPU overhead to quickly respond to transient disturbances and maintain the system's basic thermal balance, achieving an optimal balance between control accuracy and computational efficiency. It should be noted that the proportional-integral-derivative algorithm used in the low-overhead PID feedback scheduling mode is a mature existing technology in the field of industrial control, and its specific mathematical model and parameter tuning method will not be elaborated here.

[0103] Step S306 is a crucial step in pathologically diagnosing fan-induced asthma. Based on frequency domain analysis in control theory, the phase lag matching degree under the current oscillation period is calculated. The formula combines the pure lag dead time and the heat conduction time constant to calculate the total phase lag. If the current oscillation frequency results in a total phase lag of approximately 180 degrees and the matching degree falls within the critical range, it is determined to be a delay-induced strong resonance. This step determines whether intervention is necessary. It couples and verifies the apparent frequency extracted in S304 with the physical parameters analyzed in S203. Without step S306, the system might incorrectly suppress resonance caused by random fluctuations from external disturbances, leading to unnecessary performance loss; or it might remain inactive when true thermal resonance occurs because the cause cannot be identified. It acts as a firewall to prevent misjudgment, ensuring that a weighted thermal resonance state vector is generated and the damping strategy is activated only when the scheduling algorithm and thermal inertia truly resonate.

[0104] Step S4: After switching to the MPC predictive damping scheduling mode, construct and solve the dynamic weighted multi-objective optimization function based on the weighted thermal resonance state vector to generate the optimal damping scheduling command;

[0105] The specific steps in step S4 for generating the optimal damping scheduling instruction include:

[0106] Step S401: Analyze the second component in the weighted thermal resonance state vector, dynamically adjust the prediction time domain length of MPC based on the second component, extract the trajectory point corresponding to the current time from the denoised evolution trajectory, establish the discrete state evolution equation with the trajectory point corresponding to the current time, and obtain the initial prediction control decision matrix.

[0107] Extracting the second component of the weighted thermal resonance state vector As the core basis for anchored prediction in the time domain, this component physically represents the effective thermal oscillation period weighted by confidence level. When the system experiences delay-induced strong resonance, the phase lag matching degree... It will approach 1, and usually falls into The interval, at which this component is located. This directly reflects the true physical cycle of the current thermal oscillation. In order for the MPC controller to fully predict the evolution of the peaks and troughs of thermal fluctuations, the coverage ratio is used. Time-domain amplification of this component, For example, take 1.5 and set the prediction time length. Combined with sampling period Determine the final number of discrete prediction steps , This represents rounding up, utilizing the weighted properties of vectors to achieve an adaptive balance between risk and computing power. When the risk of resonance is high, the system automatically obtains a sufficiently long prediction field to effectively suppress oscillations, while when... When the risk of resonance is low, the prediction time domain is naturally shortened, thus significantly saving computational power. The current moment is extracted from the denoised evolution trajectory. The corresponding trajectory points are used as the pure prediction starting state, based on the thermal time constant. and sampling period Determined state transition matrix With input control matrix Construct discrete state evolution equations for future MPC simulations: ,in Represents the current system sampling time. For the range of values Prediction step size index, For the current moment For the future The predicted temperature value of the system's thermal state at any given time. For the time domain length, when At that time, the value is taken from the S302 noise reduction evolution trajectory. The mapping point at time is used to ensure that the prediction starts from the pure resonance trend after eliminating non-periodic divergent interference. The sequence of control inputs to be solved is applied at the predicted future time; based on this, the prediction time domain... The control input vector to be solved at all times within the time interval Stacked vertically in chronological order, the construction dimension is... Predictive control decision matrix This matrix will serve as the core optimization variable for subsequent multi-objective optimization problems, where The dimension representing the control input vector, i.e., the number of independent controllable variables in the system, is used in this invention. It consists of two parts: one is the chip power consumption control variable that adjusts computing performance, and the other is the fan speed control variable that adjusts heat dissipation capacity. , The matrix structure clearly serves subsequent optimization: the first column of the matrix is ​​directly used to calculate the performance index that maximizes computing power, the second column is used to calculate the penalty term for limiting noise, and the temperature trajectory predicted by the model by substituting the data from both columns together is used to calculate the thermal stability index, thereby realizing the direct solution of the multi-objective optimization problem.

[0108] Step S402: Construct a dynamic weighted multi-objective optimization function using the predictive control decision matrix as the decision variable. The multi-objective optimization function includes a stability term that minimizes the variance of the temperature trajectory in the prediction time domain and a performance term that maximizes the computing power throughput. A fan noise penalty factor is added, and the stability term is dynamically weighted and amplified by the resonance suppression gain coefficient modulated by the weighted thermal resonance state vector.

[0109] Constructed multi-objective optimization function It consists of three weighted parts, the first part being the stability term. Its calculation method is to predict the time domain. The discrete state evolution equation constructed in step S401 generates a future temperature trajectory based on the predictive control decision matrix. The variance or mean square error of this trajectory relative to the target temperature is calculated to characterize the system's thermal stability. The target temperature is a preset maximum safe junction temperature determined according to the chip hardware manual. The second part is the performance item. Its calculation method is based on the decision matrix. The first column contains the chip power consumption control parameters, calculated as the negative of the sum of power consumption in the prediction time domain, to represent the pursuit of maximizing computing throughput. The third part is the fan noise penalty factor. The calculation method involves using the second column element of the decision matrix ΔU, i.e., the fan speed control value, to calculate the sum of squares of the rate of change of the fan speed amplitude, which is used to limit the acoustic noise generated by drastic changes in fan speed. The specific implementation of the dynamic weighted amplification is achieved through a resonance suppression gain coefficient. The magnitude of the weighted thermal resonance state vector is modulated in real time. When the magnitude of the weighted thermal resonance state vector exceeds a preset safety threshold, it indicates that thermal resonance is about to occur in the system. At this time, the coefficient... It increases exponentially, specifically when the modulus is below a preset safety threshold. The default value is 1; if the modulus exceeds the threshold, A function is constructed using the natural constant 2.718 as the base and the difference between the current modulus and the threshold as the exponent, achieving a rapid, non-linear increase in the numerical value. The optimized function is expressed as follows: , The calibration is based on the statistical characteristics of the system's historical operating data. It involves analyzing the fluctuation range and sensitivity of each indicator, combined with static weighting coefficients determined by design requirements. Furthermore, to eliminate the influence of different physical dimensions, all three indicators are normalized before calculation to map to the [0, 1] interval. When a resonance trend is detected, the stability term... The weights are forcibly amplified, forcing the optimizer to sacrifice computational performance in exchange for the thermal stability of the system.

[0110] Step S403: Solve the multi-objective optimization function using the rolling time-domain optimization algorithm, and output the optimal damping scheduling command under the premise of satisfying the core maximum safe temperature constraint.

[0111] Multi-objective optimization function The problem is transformed into a standard quadratic programming or nonlinear programming problem for solution. During the solution process, strict adherence to physical constraints is essential. These constraints include ensuring that the predicted chip junction temperature at any given time within the prediction time domain does not exceed the core's maximum safe temperature (e.g., 95°C), and that control variables (fan speed, operating frequency) remain within the hardware-allowed physical limits. The rolling time-domain optimization algorithm optimizes the chip junction temperature at the current sampling time. 2. The solver calculates the future The optimal control sequence at each time step is the optimal predictive control decision matrix. However, the system only selects the first control element in this sequence as the actual command to be issued to the actuator. The first control element is the current sampling time. The control quantity. The optimal damping scheduling instruction output is the first control element selected above. This instruction contains specific target fan speed values ​​and chip frequency limit values. Its physical effect is equivalent to applying active damping to the thermal system, which can effectively suppress temperature overshoot and oscillation caused by thermal inertia, thereby maximizing computing power release while ensuring that overheating and frequency reduction are not triggered. At the next sampling time. The system will repeat steps S401 to S403 above, using new measurement data to refresh and optimize the problem, and achieve closed-loop control.

[0112] Step S403 employs a rolling time-domain optimization algorithm to solve a multi-objective optimization function that includes stability, performance, and fan noise terms, outputting the optimal control quantity at the current moment. This is the final execution method for solving the problem. After detecting a limit cycle trajectory in step S303 and confirming delay-induced strong resonance in step S306, S403 forcibly amplifies the weight of the stability term in the optimization function, effectively applying active damping to the system. It no longer presents a choice between overheating and underheating, but calculates a precise intermediate value, smoothly suppressing fluctuations in fan speed and core temperature. Without this step, all perception and analysis cannot be translated into actual control actions, and the physical phenomenon of fan asthma cannot be eliminated.

[0113] Example 2

[0114] like Figure 2 As shown, the server energy-saving system based on a hysteresis-compensated dynamic thermal impedance model includes a time series graph module, a model generation module, a resonance analysis module, a scheduling instruction module, an independent edge inference module, and a lightweight model storage module. Specifically:

[0115] The time series graph module is used to collect the underlying operating data of the server hardware in real time, map the operating data to a unified time axis, and construct a multi-dimensional heat-computing power coupled time series graph;

[0116] Based on the multidimensional thermal-computational coupling time series graph, the model generation module calculates the heat conduction time constant under the current hardware environment and establishes a hysteresis-compensated dynamic thermal impedance model based on the heat conduction time constant.

[0117] The resonance analysis module constructs a thermodynamic phase space based on the operating data, analyzes the noise reduction evolution trajectory in conjunction with the hysteresis-compensated dynamic thermal impedance model, and performs delay-induced strong resonance analysis based on the noise reduction evolution trajectory to generate a weighted thermal resonance state vector.

[0118] The scheduling instruction module is based on the weighted thermal resonance state vector, activates the adaptive gain MPC predictive damping scheduling strategy, constructs and solves the dynamic weighted multi-objective optimization function, and generates the optimal damping scheduling instruction.

[0119] The independent edge inference module, independent of the server's main processor, serves as the physical execution carrier for the aforementioned modules, acquiring data and executing computational tasks via the sideband bus.

[0120] The independent edge inference module is a hardware computing unit independent of the server's main processor. Serving as the physical execution carrier for steps S103 to S4 of the aforementioned method, it is physically deployed within an independent ARM core of the server motherboard's baseboard management controller, or constructed using an independent low-power microcontroller unit and a dedicated neural processing unit. Logically, this module does not pass through the server operating system kernel. Specifically, it establishes a hardware communication link directly with the temperature sensors and fan controllers on the motherboard via sideband buses such as I2C, PECI, or USB. It is responsible for real-time data acquisition, time-series graph construction, and model inference calculations. The introduction of this independent hardware enables out-of-band thermal management, meaning that complex neural network inference and PID / MPC calculation tasks are completely separated from the main CPU, without occupying any of the main CPU's computing resources. This avoids the risk of deadlock due to CPU resource contention during high server load and overheating. This architecture, utilizing an independent microcontroller or BMC for out-of-band management and hardware monitoring, is a common hardware design scheme in the server hardware architecture field. As a common existing technology, it will not be elaborated upon here.

[0121] The lightweight model storage module stores the lightweight neural network parameters after pruning and quantization, providing parameter support for the time series graph module.

[0122] The lightweight model storage module is a firmware unit integrated into the non-volatile storage space of the aforementioned independent edge inference module. The process of mapping the time-series graph neural network into a topology as described in step S103 specifically involves the independent edge inference module calling the parameters in this storage module for computation. Internally, it stores lightweight neural network parameters trained with offline big data and processed through pruning and quantization. Specifically, the 32-bit floating-point weight parameters in the original model are quantized into 8-bit integers, compressing the model size and computational load to less than 5% of the original while retaining more than 95% of the model's accuracy. This allows complex nonlinear heat conduction deductions to run in real time at millisecond speeds on MCU hardware with limited computing power. This ensures that even under extreme conditions such as the main CPU overheating and crashing or the operating system crashing, the module can still independently and continuously monitor and take over heat dissipation control, ensuring the physical safety of the server. This method of pruning and quantizing neural network models to adapt to edge device deployment is a common data processing technique in the fields of artificial intelligence and edge computing. As a common existing technology, it will not be elaborated on here.

[0123] It should be understood that although the steps in the flowcharts of the various embodiments of the present invention are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the various embodiments may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least a portion of the sub-steps or stages of other steps.

[0124] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments described above. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0125] The foregoing description is illustrative of the invention and should not be construed as limiting it. Although several exemplary embodiments of the invention have been described, those skilled in the art will readily understand that many modifications can be made to the exemplary embodiments without departing from the novel teachings and advantages of the invention. Therefore, all such modifications are intended to be included within the defined scope of the invention. It should be understood that the foregoing description is illustrative of the invention and should not be construed as limiting it to the specific embodiments disclosed, and modifications to the disclosed embodiments and other embodiments are intended to be included within the appended scope of protection. The invention is defined by its scope of protection and its equivalents.

Claims

1. A server energy-saving method based on a hysteresis-compensated dynamic thermal impedance model, characterized in that, Includes the following steps: Real-time acquisition of server hardware underlying operation data, mapping of the operation data to a unified time axis, and construction of a multi-dimensional thermal-computing power coupling time series graph; Based on the multidimensional thermal-computing coupling time series diagram, the heat conduction time constant under the current hardware environment is calculated, and a hysteresis-compensated dynamic thermal impedance model is established based on the heat conduction time constant. The steps for establishing a hysteresis-compensated dynamic thermal impedance model are as follows: A three-dimensional heat conduction geometry is constructed based on physical packaging parameters, and the discrete sampling data of the sensor is mapped to the boundary thermal potential energy constraint of the geometry surface to construct a reverse heat flow cone pointing towards the chip core. An isothermal manifold surface cluster is generated within the reverse heat flow cone. The path backtracking is performed using the geometric axiom that the heat flow gradient vector is orthogonal to the isothermal surface, and the core true temperature field is obtained by superimposing the energy flux conservation constraint. Utilizing the thermal conduction time constant Pure time delay dead zone parameter and steady-state gain Construct a first-order inertial element model with pure time delay: ,in This represents the actual temperature of the chip core. This value is extracted from the actual temperature field of the chip core, specifically from the temperature data corresponding to the core node location within the temperature field obtained by cropping and superimposing the data from the chip core surface to the heat spreader layer. It was delayed The input power consumption over time, i.e., the corrected computing power power density, This refers to the environmental temperature disturbance term that drifts slowly over time. A Kalman filter correction term is introduced to compensate for ambient temperature drift, generating a hysteresis-compensated dynamic thermal impedance model. A thermodynamic phase space is constructed based on the operating data. The noise reduction evolution trajectory is analyzed in conjunction with the hysteresis compensation dynamic thermal impedance model. Delay-induced strong resonance analysis is performed based on the noise reduction evolution trajectory to generate a weighted thermal resonance state vector and switch the MPC predictive damping scheduling mode. After switching to the MPC predictive damping scheduling mode, a dynamic weighted multi-objective optimization function is constructed and solved based on the weighted thermal resonance state vector to generate the optimal damping scheduling command.

2. The server energy-saving method based on a hysteresis-compensated dynamic thermal impedance model according to claim 1, characterized in that, The analytical steps for constructing the multidimensional heat-computing power coupled time series graph are as follows: The core real temperature field and fan PWM duty cycle are obtained through the in-band sensor interface, and the instruction set intensity factor and cache miss rate are extracted simultaneously through the operating system performance counter to form a heterogeneous multidimensional state flow. Sliding window timing analysis is performed on the heterogeneous multidimensional state flow. The heat dissipation flux is calculated based on the spatial gradient of the core real temperature field. The computing power power density per unit time is deduced by combining the instruction set strength factor and cache miss rate, and a heat-power-heat dissipation coupled feature vector is generated. The thermal-power-heat dissipation coupling feature vector is mapped into a topological structure using a time-series graph neural network, where nodes represent the physical core and edge weights represent the thermal conduction correlation strength, generating a multidimensional thermal-computing power coupling time-series graph.

3. The server energy-saving method based on a hysteresis-compensated dynamic thermal impedance model according to claim 2, characterized in that, The analysis steps for the heat conduction time constant, pure time delay parameter, and steady-state gain are as follows: Step response segments are extracted from the multidimensional thermal-computing coupling time series graph, and discrete transfer function coefficients are identified using the recursive least squares method with forgetting factor. Based on the discrete transfer function coefficients, the heat conduction time constant, the pure time delay dead zone parameter, and the steady-state gain of the system are analyzed.

4. The server energy-saving method based on a hysteresis-compensated dynamic thermal impedance model according to claim 1, characterized in that, The analytical steps for generating the weighted thermal resonance state vector are as follows: Using the pure lag dead time parameter and the heat conduction time constant as physical constraints, the oscillation frequency feature tensor is mapped to the lag-period coupling discriminant function to calculate the phase lag matching degree. If the phase lag matching degree falls into the preset critical resonance range, the current state is determined to be a delay-induced strong resonance, a weighted thermal resonance state vector is generated, and the system switches to the MPC predictive damping scheduling mode. Otherwise, it is determined to be a transient load disturbance and the low-overhead PID feedback scheduling mode is activated.

5. The server energy-saving method based on a hysteresis-compensated dynamic thermal impedance model according to claim 4, characterized in that, The analysis steps for the oscillation frequency characteristic tensor are as follows: Based on the limit cycle trajectory, the time regression period of the noise reduction evolution trajectory on the Poincaré section is extracted, and combined with the thermodynamic phase space dimension information, an oscillation frequency feature tensor is constructed.

6. The server energy-saving method based on a hysteresis-compensated dynamic thermal impedance model according to claim 5, characterized in that, The analysis steps for the limit cycle trajectory are as follows: A Poincaré section is set in the thermodynamic phase space, and the regression characteristics of the noise reduction evolution trajectory crossing the section are detected. If the Euclidean distance between the crossing points is less than a preset convergence threshold multiple times, the trajectory is determined to form a limit cycle trajectory and is confirmed to be in a thermal oscillation resonance state.

7. The server energy-saving method based on a hysteresis-compensated dynamic thermal impedance model according to claim 6, characterized in that, The analysis steps for the noise reduction evolution trajectory are as follows: The collected instantaneous temperature data are subjected to first-order and second-order difference operations to extract dynamic feature vectors and construct a thermodynamic phase space. The dynamic feature vector is mapped to discrete trajectory points in the thermodynamic phase space. The hysteresis-compensated dynamic thermal impedance model is used to eliminate non-periodic trajectory divergence. The discrete trajectory points are then corrected and reconstructed to obtain the noise-reduced evolution trajectory.

8. The server energy-saving method based on a hysteresis-compensated dynamic thermal impedance model according to claim 1, characterized in that, The analysis steps for generating the optimal damping scheduling command are as follows: The second component in the weighted thermal resonance state vector is analyzed, and the prediction time domain length of MPC is dynamically adjusted based on the second component. The trajectory point corresponding to the current time is extracted from the denoised evolution trajectory, and the discrete state evolution equation is established with the trajectory point corresponding to the current time to obtain the initial prediction control decision matrix. A dynamic weighted multi-objective optimization function is constructed using the predictive control decision matrix as the decision variable. The multi-objective optimization function includes a stability term that minimizes the variance of the temperature trajectory in the prediction time domain and a performance term that maximizes the computing power throughput. A fan noise penalty factor is added, and the stability term is dynamically weighted and amplified by the resonance suppression gain coefficient modulated by the weighted thermal resonance state vector. The multi-objective optimization function is solved using a rolling time-domain optimization algorithm, and the optimal damping scheduling command is output under the premise of satisfying the core's highest safe temperature constraint.

9. A server energy-saving system based on a hysteresis-compensated dynamic thermal impedance model, characterized in that, The server energy-saving method based on the hysteresis-compensated dynamic thermal impedance model as described in any one of claims 1-8 includes: The time series graph module is used to collect the underlying operating data of the server hardware in real time, map the operating data to a unified time axis, and construct a multi-dimensional heat-computing power coupled time series graph; Based on the multidimensional thermal-computational coupling time series graph, the model generation module calculates the heat conduction time constant under the current hardware environment and establishes a hysteresis-compensated dynamic thermal impedance model based on the heat conduction time constant. The resonance analysis module constructs a thermodynamic phase space based on the operating data, analyzes the noise reduction evolution trajectory in combination with the hysteresis-compensated dynamic thermal impedance model, and performs delay-induced strong resonance analysis based on the noise reduction evolution trajectory to generate a weighted thermal resonance state vector and switch the MPC predictive damping scheduling mode. After switching to the MPC predictive damping scheduling mode, the scheduling instruction module constructs and solves a dynamic weighted multi-objective optimization function based on the weighted thermal resonance state vector to generate the optimal damping scheduling instruction. The independent edge inference module, independent of the server's main processor, serves as the physical execution carrier for the aforementioned modules, acquiring data and executing computational tasks via the sideband bus. The lightweight model storage module stores the lightweight neural network parameters after pruning and quantization, providing parameter support for the time series graph module.

Images